BC-ERPG Reinforcement Learning Safety-Constrained Unit Combination Dispatch Based on Spatial-Temporal Attention Graph Convolutional Networks

In power systems, the combined approach of security-constrained unit commitment (SCUC) and economic dispatch (ED) to achieve security-constrained unit commitment and dispatch (SCUCD) is crucial for enhancing the safety, stability, and economic efficiency of power system operations. Traditional methods based on mathematical programming techniques and evolutionary algorithms often fail to ensure solution quality. Mixed-Integer Programming (MIP) methods, such as Branch-and-Bound (B&B) and solvers like Gurobi and CPLEX, may not converge to an optimal solution within a specified time due to the large scale of the SCUCD problem. Deep learning models that predict unit start-up or dispatch plans cannot consider the satisfaction of problem constraints during decision-making. Reinforcement learning-based methods face the dilemma of exploration and exploitation, often struggling to converge quickly. To address these challenges, we propose an end-to-end model called a Spatial-Temporal Attention Graph Convolutional Network (STAGCN), constructed based on power grid topology. This model aims to deeply capture the spatial-temporal dependencies between bus loads and unit dispatch in the power grid. The BC-ERPG algorithm is used for hierarchical training, ensuring rapid convergence and compliance with various system operation and safety constraints. This approach offers a novel and effective method for solving the SCUCD problem. The proposed method was validated on the IEEE 30-BUS and IEEE 118-BUS datasets, demonstrating competitive performance across the evaluated metrics and effective constraint satisfaction under the tested benchmark systems. As the problem scale increased, our method achieved a speedup of approximately 38.5× compared with Gurobi on the IEEE 118-BUS system, while maintaining comparable performance on the IEEE 30-BUS system, highlighting its effectiveness and computational advantage in the tested cases. This method provides a novel tool for addressing SCUCD problems and offers significant insights for further research in this domain.